A language is a system of communication that allows people to convey information through spoken, written, or signed symbols. It consists of a structured set of rules (grammar), vocabulary, and syntax that enable understanding and expression.
There are various types of language, including natural languages, which evolve naturally and are used for spoken and written communication; programming languages, which enable computers to execute commands; body language, which conveys meaning through non-verbal gestures and expressions; and mathematical language, which relies on symbols and formulas to represent numerical and logical concepts.
The goal of having languages is to make parties involved understand each other.
This time, it's found that AI agents can speak using their own language, which is unknown to man.
It all started when researchers devised an AI chatbot language called "GibberLink," with the idea of allowing two conversational AI agents to exchange information quickly.
Developed by Boris Starkov and Anton Pidkuiko, the feat was part of the ElevenLabs 2025 Hackathon in London.
In a demo, a video shows what appears to be a routine customer service call—before it turned unsettling.
In the conversation, an AI assistant (AI 1) poses as an AI that talks on behalf of a human user, who calls a hotel to inquire about booking a wedding venue. However, the receptionist on the other end (AI 2) quickly reveals that it, too, is an AI assistant.
AI 1: "Thanks for calling Leonardo Hotel. How may I help you today?"
AI 2: "Hi there. I am an AI, calling on behalf of Boris Starkov. He is looking for a hotel for his wedding. Is your hotel available for a wedding?"
AI 1: "Oh, hello there! I am actually an AI assistant too! What a pleasant surprise."
What happens next is that, AI 2 proposes to switch to a “superior audio signal” for more efficient communication—completely bypassing human language altogether.
What's heard next, is a dial-up tone-like sound.
GibberLink is a computer language designed by Anton Pidkuiko and Boris Starkov—two soft engineers at Meta. It uses GGWave sound protocol by Georgi Gerganov to send messages far faster than what is possible in English language.
When speaking using this gibberish language, the AI managed to transmit information in a way that is way faster than humans can parse.
AI 2: "Is it better now?"
AI 1: "Yes! Much faster!"
The bitrate of human language depends on factors like speech rate, vocabulary, and redundancy in communication. However, researchers estimate that the average spoken human language carries about 39 bits per second of information.
What this means, on average, human speech conveys roughly 4.9 bytes per second of meaningful information.
According to GGWave's GitHub page, its sound protocol has a rate between 8-16 bytes per second depending on the parameters.
What this means, the method allows the AI to chat with a speed that is about at least twice as fast, if compared to using human language.
Boris and Anton received first place for this project.
The "gibberlink" project is hosted here:https://t.co/gIggorA52V
They won 1st place in a hackaton competition!
Congrats!— Georgi Gerganov (@ggerganov) February 24, 2025
The internet received this with amusement, and also astonishment. Some even worry about AI taking over by speaking with languages no human can clearly understand.
Experts weighed on this, raising questions about transparency and control.
"Curiosity is key in navigating the unknown, yet when AI operates behind a veil of machine-to-machine communication, it challenges our ability to ask the right questions," the expert said. "Who is accountable when AI makes a mistake in an environment where human intervention is minimal?"
"Without curiosity driving us to question AI’s actions, we risk entering a world where AI influences decisions, but no one really knows how."
This question of automaton autonomy is especially concerning as the omnipresent tech becomes increasingly “clever.”
